pi-coralbricks-provider 1.0.16 → 1.0.18

package/README.md

@@ -2,7 +2,7 @@
 
 # 🪸 pi-coralbricks-provider
 
-**GLM 5.3, GLM 5.3 Flash, Kimi K3 & GPT-OSS 120B through [Coral Bricks](https://www.coralbricks.ai)**
+**GLM 5.3, GLM 5.3 Flash & DeepSeek V4.1 Flash through [Coral Bricks](https://www.coralbricks.ai)**
 
 _A [pi](https://github.com/earendil-works/pi-coding-agent) provider extension for Coral's OpenAI-compatible inference gateway — up to **1M context** on open models._
 
@@ -17,13 +17,13 @@ _A [pi](https://github.com/earendil-works/pi-coding-agent) provider extension fo
 
 ## Features
 
-- **4 reasoning models** from Coral's live catalog — GLM 5.3 FP4, GLM 5.3 Flash, Kimi K3, and GPT-OSS 120B
-- **1M token context** on GLM and Kimi, with vision (image input) on Kimi K3 and GLM 5.3 Flash
+- **3 reasoning models** from Coral's live catalog — GLM 5.3 FP4, GLM 5.3 Flash, and DeepSeek V4.1 Flash
+- **1M token context** on GLM and DeepSeek, with vision (image input) on GLM 5.3 Flash and DeepSeek V4.1 Flash
 - **OpenAI-compatible API** — standard `/v1/chat/completions`, streaming, and tool calling
-- **Per-family thinking levels** — zai-style `thinking` control for GLM (including a *real* off switch), `reasoning_effort` for Kimi K3 and GPT-OSS
-- **Accurate cost tracking** — pricing mirrors Coral's published rates, and cached reads are **$0 on every model**, so pi's computed cost matches Coral's own `usage.cost` to the token
+- **Per-family thinking levels** — zai-style `thinking` control for GLM (including a *real* off switch), `reasoning_effort` for DeepSeek V4.1 Flash
+- **Accurate cost tracking** — input, cache-write and output rates mirror Coral's [published pricing](https://www.coralbricks.ai/pricing), and cached reads are **$0 on every model**
 - **Self-healing model sync** — stale-while-revalidate from the authenticated `/v1/models` (or the unauthenticated [public catalog](https://www.coralbricks.ai/api/public/models) before auth), hot-swapped at session start
-- **synbad-validated** — [synbad](https://github.com/synthetic-lab/synbad) tool-calling and reasoning-parsing evals pass 13/13 in every run on GLM 5.3 and Kimi K3 in both unary and streaming modes
+- **synbad-validated** — [synbad](https://github.com/synthetic-lab/synbad) tool-calling and reasoning-parsing evals pass 13/13 in every run on GLM 5.3 in both unary and streaming modes
 
 ## Installation
 
@@ -78,14 +78,13 @@ pi
 
 ## Available Models
 
-| Model | Context | Vision | Reasoning | Input $/M | Cache Read $/M | Output $/M |
-|-------|---------|--------|-----------|-----------|-----------------|------------|
-| DeepSeek V4.1 Flash | 1.0M | ❌ | ❌ | $0.30 | — | $1.20 |
-| GLM 5.3 Flash | 1.0M | ✅ | ✅ | $0.15 | — | $0.50 |
-| GLM 5.3 FP4 | 1.0M | ❌ | ✅ | $1.12 | — | $4.40 |
-| GPT-OSS 120B | 131K | ❌ | ✅ | $0.12 | — | $0.60 |
+| Model | Context | Vision | Reasoning | Input $/M | Cache Read $/M | Cache Write $/M | Output $/M |
+|-------|---------|--------|-----------|-----------|-----------------|------------------|------------|
+| DeepSeek V4.1 Flash | 1.0M | ✅ | ✅ | $0.30 | — | $0.09 | $1.20 |
+| GLM 5.3 Flash | 1.0M | ✅ | ✅ | $0.15 | — | $0.23 | $0.50 |
+| GLM 5.3 FP4 | 1.0M | ❌ | ✅ | $1.12 | — | $1.68 | $4.40 |
 
-*Costs are per million tokens. Cache Read shows — because Coral bills cached input at **$0** on every model. Prices subject to change — check [Coral's live catalog](https://www.coralbricks.ai/api/public/models).*
+*Costs are per million tokens. Cache Read shows — because Coral bills cached input at **$0** on every model. Prompt tokens Coral has not cached yet are billed once at the Cache Write rate, in place of the Input rate. Prices subject to change — check [Coral's live catalog](https://www.coralbricks.ai/api/public/models).*
 
 ## Usage
 
@@ -98,10 +97,10 @@ After loading the extension, use the `/model` command in pi to select your prefe
 Or start pi directly with a Coral model:
 
 ```bash
-pi -e /path/to/pi-coralbricks-provider --model coralbricks/kimi-k3:high
+pi -e /path/to/pi-coralbricks-provider --model coralbricks/deepseek-v4.1-flash-fast-fp4:high
 ```
 
-Thinking levels attach to the model id with `:<level>` — e.g. `:low`, `:high`, `:max`, or `:off` (GLM only).
+Thinking levels attach to the model id with `:<level>` — e.g. `:low`, `:high`, `:max`, or `:off` (GLM and DeepSeek).
 
 ## Authentication
 
@@ -131,12 +130,10 @@ Verified against the live gateway:
 |-------|--------|-----|-----|--------|------|-----|
 | GLM 5.3 FP4 | `thinking: {type}` + `reasoning_effort` | ✅ | ✅ | — | ✅ | ✅ |
 | GLM 5.3 Flash | `thinking: {type}` + `reasoning_effort` | ✅ | ✅ | — | ✅ | ✅ |
-| Kimi K3 | `reasoning_effort` | — | ✅ | — | ✅ | ✅ |
-| GPT-OSS 120B | `reasoning_effort` | — | ✅ | ✅ | ✅ | — |
+| DeepSeek V4.1 Flash | `reasoning_effort` | ✅ | ✅ | — | ✅ | ✅ |
 
 - **GLM** accepts zai-style `thinking: {type: "disabled"}` on Coral — pi's *off* level turns thinking off, though a short preamble of a few dozen tokens can still appear (the upstream Z.ai API has no off at all, so this differs from the canonical Z.ai map).
-- **Kimi K3** always thinks: `reasoning_effort: "none"` is accepted but doesn't disable reasoning, so *off* is intentionally not offered. Assistant replays include `reasoning_content`.
-- **GPT-OSS** exposes the standard low/medium/high reasoning efforts; reasoning arrives in `reasoning_content`.
+- **DeepSeek V4.1 Flash** reasons only when asked: a request without `reasoning_effort`, or with `"none"`, gets no reasoning, which is what pi's *off* sends. `thinking: {type}` has no effect on this model.
 - Coral streams a duplicate `reasoning` field alongside `reasoning_content`; pi dedupes these automatically.
 
 ## Compat Settings
@@ -146,30 +143,29 @@ Coral's gateway follows the OpenAI Chat Completions API:
 - **`supportsStore: false`** / **`supportsDeveloperRole: false`** — all models; Coral serves open models on the classic roles.
 - **`maxTokensField: "max_tokens"`** — all models.
 - **`thinkingFormat: "zai"`** — GLM 5.3: `thinking: {type: "enabled"|"disabled"}` toggles reasoning, `reasoning_effort` picks the depth.
-- **`thinkingFormat: "openai"`** — Kimi K3 and GPT-OSS 120B: `reasoning_effort` drives thinking depth.
-- **`supportsStrictMode: false`** — Kimi K3 (no strict JSON-schema tool definitions).
-- **`requiresReasoningContentOnAssistantMessages: true`** — Kimi K3.
+- **`thinkingFormat: "openai"`** — DeepSeek V4.1 Flash: `reasoning_effort` drives thinking depth.
 
 ### Patch Overrides & Custom Models
 
-- **`patch.json`** — per-model overrides applied on top of `models.json` (reasoning flags, pricing corrections, compat settings, thinking level maps). Currently carries the GLM 5.3 Flash entry: reasoning, image input, its thinking level map and compat settings.
+- **`patch.json`** — per-model overrides applied on top of `models.json` (reasoning flags, pricing corrections, compat settings, thinking level maps). Currently carries two entries: GLM 5.3 Flash (reasoning, image input, its thinking level map and compat settings) and DeepSeek V4.1 Flash (reasoning, its thinking level map and compat settings).
 - **`custom-models.json`** — full model definitions for models Coral doesn't list. Merged after patch.
 
 Merge order: `[live|cache|embedded] → patch.json → custom-models.json`
 
 ## Inference-Quality Testing
 
-Validated with [synbad](https://github.com/synthetic-lab/synbad) — Synthetic's tool-calling and reasoning-parsing eval suite for LLM inference providers (current 13-eval suite, `--reasoning-effort high`, five `--count 1` runs per mode on 2026-09-15; a cell shows the evals that passed in every run):
+Validated with [synbad](https://github.com/synthetic-lab/synbad) — Synthetic's tool-calling and reasoning-parsing eval suite for LLM inference providers (current 13-eval suite, `--reasoning-effort high`, five `--count 1` runs per mode on 2026-09-15, DeepSeek on 2026-09-22; a cell shows the evals that passed in every run):
 
 | Model | Unary | Stream | Notes |
 |-------|-------|--------|-------|
 | GLM 5.3 FP4 | 13/13 ✅ | 13/13 ✅ | 5/5 runs |
 | GLM 5.3 Flash | 12/13 ⚠️ | 12/13 ⚠️ | `reasoning/reasoning-parsing` passes 4 of 5 runs; see below |
-| Kimi K3 | 13/13 ✅ | 13/13 ✅ | 5/5 runs |
-| GPT-OSS 120B | 12/13 ⚠️ | 12/13 ⚠️ | `tools/parallel-tool` 0 of 5 runs; see below |
+| GPT-OSS 120B (retired) | 12/13 ⚠️ | 12/13 ⚠️ | `tools/parallel-tool` 0 of 5 runs, before Coral retired the model |
+| DeepSeek V4.1 Flash | 12/13 ⚠️ | 12/13 ⚠️ | `tools/octo-list-no-optional-args` passes 1 of 5 runs; see below |
+| Kimi K3 (retired) | 13/13 ✅ | 13/13 ✅ | 5/5 runs, before Coral retired the model |
 | GLM 5.2 FP4 (retired) | 15/15 ✅ | 15/15 ✅ | 15-eval suite, single run, before Coral retired the model |
 
-The two misses are model-side, not transport: gpt-oss answers "Paris and London" with one call after another instead of two parallel calls even when `parallel_tool_calls: true`, in every run; and on the trivial prompt `reasoning-parsing` uses, GLM 5.3 Flash answers without a think block about one run in five, in both modes, with the answer itself unaffected.
+The misses are model-side, not transport: gpt-oss answered "Paris and London" with one call after another instead of two parallel calls even when `parallel_tool_calls: true`, in every run; on the trivial prompt `reasoning-parsing` uses, GLM 5.3 Flash answers without a think block about one run in five, in both modes, with the answer itself unaffected; and asked to "call the list tool with no args", DeepSeek V4.1 Flash usually fills in the optional `dirPath` as `"."`.
 
 Reproduce:
 

package/deprecated-models.json

@@ -1,38 +1 @@
-{
-  "kimi-k3": {
-    "id": "kimi-k3",
-    "name": "Kimi K3",
-    "reasoning": true,
-    "input": [
-      "text",
-      "image"
-    ],
-    "cost": {
-      "input": 3,
-      "output": 15,
-      "cacheRead": 0,
-      "cacheWrite": 0
-    },
-    "contextWindow": 1048576,
-    "maxTokens": 131072,
-    "thinkingLevelMap": {
-      "off": null,
-      "minimal": null,
-      "low": "low",
-      "medium": null,
-      "high": "high",
-      "xhigh": null,
-      "max": "max"
-    },
-    "compat": {
-      "supportsStore": false,
-      "supportsDeveloperRole": false,
-      "supportsReasoningEffort": true,
-      "maxTokensField": "max_tokens",
-      "supportsStrictMode": false,
-      "thinkingFormat": "openai",
-      "requiresReasoningContentOnAssistantMessages": true
-    },
-    "deprecatedAt": "2026-09-21T00:00:07.264Z"
-  }
-}
+{}

package/index.ts

@@ -224,7 +224,8 @@ function baseCompat() {
 /**
  * Transform a Coral /v1/models row. Coral reports per-million USD pricing
  * (input_per_m / output_per_m / cached_input_per_m — cached reads are $0 on
- * every model), context_length, and supports_image_input.
+ * every model — and cache_write_per_m, the rate for prompt tokens written to
+ * the cache), context_length, and supports_image_input.
  */
 function transformApiModel(apiModel: any): JsonModel | null {
   if (!apiModel?.id) return null;
@@ -238,7 +239,7 @@ function transformApiModel(apiModel: any): JsonModel | null {
       input: toNumber(pricing.input_per_m),
       output: toNumber(pricing.output_per_m),
       cacheRead: toNumber(pricing.cached_input_per_m),
-      cacheWrite: 0,
+      cacheWrite: toNumber(pricing.cache_write_per_m),
     },
     contextWindow: parseContextWindow(apiModel.context_length),
     maxTokens: DEFAULT_MAX_TOKENS[apiModel.id] ?? FALLBACK_MAX_TOKENS,
@@ -252,7 +253,9 @@ function transformApiModel(apiModel: any): JsonModel | null {
  * (https://www.coralbricks.ai/api/public/models) — the no-auth mirror of the
  * gateway control plane. Fields: slug, name, contextWindow ("1M"),
  * inputPerM / outputPerM. Coral's own prices are authoritative here; the
- * parity/field vendor comparison fields are ignored.
+ * parity/field vendor comparison fields are ignored. The catalog carries no
+ * cache-write rate, so cacheWrite stays 0 and mergeWithEmbedded keeps the
+ * embedded rate.
  */
 function transformCatalogModel(entry: any): JsonModel | null {
   if (!entry?.slug) return null;
@@ -347,7 +350,8 @@ function mergeWithEmbedded(liveModels: JsonModel[], embeddedModels: JsonModel[])
       // Self-heal: live API pricing is authoritative field-by-field. Prefer the
       // live cost when the API reports it (non-zero); fall back to embedded when
       // the API is silent (0). Coral reports cached_input_per_m as 0 because
-      // cached reads are free, which matches the curated cacheRead: 0.
+      // cached reads are free, which matches the curated cacheRead: 0. The
+      // public catalog has no cache-write rate, so its 0 keeps the embedded one.
       // Curation (reasoning/input/compat/name/thinkingLevelMap/maxTokens) still wins via ...embedded.
       result.push({
         ...liveModel,

package/models.json

@@ -4,13 +4,14 @@
     "name": "DeepSeek V4.1 Flash",
     "reasoning": false,
     "input": [
-      "text"
+      "text",
+      "image"
     ],
     "cost": {
       "input": 0.3,
       "output": 1.2,
       "cacheRead": 0,
-      "cacheWrite": 0
+      "cacheWrite": 0.09
     },
     "contextWindow": 1048576,
     "maxTokens": 32768,
@@ -32,7 +33,7 @@
       "input": 0.15,
       "output": 0.5,
       "cacheRead": 0,
-      "cacheWrite": 0
+      "cacheWrite": 0.23
     },
     "contextWindow": 1048576,
     "maxTokens": 32768,
@@ -53,7 +54,7 @@
       "input": 1.12,
       "output": 4.4,
       "cacheRead": 0,
-      "cacheWrite": 0
+      "cacheWrite": 1.68
     },
     "contextWindow": 1048576,
     "maxTokens": 131072,
@@ -73,37 +74,5 @@
       "maxTokensField": "max_tokens",
       "thinkingFormat": "zai"
     }
-  },
-  {
-    "id": "gpt-oss-120b",
-    "name": "GPT-OSS 120B",
-    "reasoning": true,
-    "input": [
-      "text"
-    ],
-    "cost": {
-      "input": 0.12,
-      "output": 0.6,
-      "cacheRead": 0,
-      "cacheWrite": 0
-    },
-    "contextWindow": 131072,
-    "maxTokens": 40960,
-    "thinkingLevelMap": {
-      "off": null,
-      "minimal": null,
-      "low": "low",
-      "medium": "medium",
-      "high": "high",
-      "xhigh": null,
-      "max": null
-    },
-    "compat": {
-      "supportsStore": false,
-      "supportsDeveloperRole": false,
-      "supportsReasoningEffort": true,
-      "maxTokensField": "max_tokens",
-      "thinkingFormat": "openai"
-    }
   }
 ]

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "pi-coralbricks-provider",
-  "version": "1.0.16",
-  "description": "CoralBricks provider extension for pi - Access GLM 5.3, GLM 5.3 Flash, Kimi K3, and GPT-OSS 120B through the Coral Inference API",
+  "version": "1.0.18",
+  "description": "CoralBricks provider extension for pi - Access GLM 5.3, GLM 5.3 Flash, and DeepSeek V4.1 Flash through the Coral Inference API",
   "author": "monotykamary",
   "homepage": "https://github.com/monotykamary/pi-coralbricks-provider#readme",
   "repository": {
@@ -13,6 +13,13 @@
   },
   "type": "module",
   "main": "index.ts",
+  "scripts": {
+    "clean": "echo 'nothing to clean'",
+    "build": "echo 'nothing to build'",
+    "check": "tsc --noEmit && vitest run",
+    "test": "vitest run",
+    "update-models": "node scripts/update-models.js"
+  },
   "keywords": [
     "pi-package",
     "pi",
@@ -23,8 +30,8 @@
     "ai",
     "llm",
     "glm",
-    "kimi",
-    "gpt-oss"
+    "deepseek",
+    "deepseek"
   ],
   "license": "MIT",
   "files": [
@@ -50,12 +57,5 @@
   "dependencies": {
     "@earendil-works/pi-ai": "0.87.0",
     "@earendil-works/pi-coding-agent": "0.87.0"
-  },
-  "scripts": {
-    "clean": "echo 'nothing to clean'",
-    "build": "echo 'nothing to build'",
-    "check": "tsc --noEmit && vitest run",
-    "test": "vitest run",
-    "update-models": "node scripts/update-models.js"
   }
-}
+}

package/patch.json

@@ -1,4 +1,20 @@
 {
+  "deepseek-v4.1-flash-fast-fp4": {
+    "reasoning": true,
+    "thinkingLevelMap": {
+      "off": "none",
+      "minimal": null,
+      "low": "low",
+      "medium": null,
+      "high": "high",
+      "xhigh": null,
+      "max": "max"
+    },
+    "compat": {
+      "supportsReasoningEffort": true,
+      "thinkingFormat": "openai"
+    }
+  },
   "glm-5.3-flash-fp4": {
     "reasoning": true,
     "input": ["text", "image"],

package/scripts/update-models.js

@@ -136,6 +136,7 @@ function resolveApiKey() {
 }
 
 const MODELS_API_URL = 'https://inference.coralbricks.ai/v1/models';
+const CHAT_COMPLETIONS_URL = 'https://inference.coralbricks.ai/v1/chat/completions';
 const PUBLIC_CATALOG_URL = 'https://www.coralbricks.ai/api/public/models';
 const MODELS_JSON_PATH = path.join(__dirname, '..', 'models.json');
 const PATCH_JSON_PATH = path.join(__dirname, '..', 'patch.json');
@@ -249,6 +250,7 @@ function transformApiModel(apiModel, existingModelsMap) {
     if (typeof pricing.input_per_m === 'number') existing.cost.input = pricing.input_per_m;
     if (typeof pricing.output_per_m === 'number') existing.cost.output = pricing.output_per_m;
     existing.cost.cacheRead = toNumber(pricing.cached_input_per_m);
+    if (typeof pricing.cache_write_per_m === 'number') existing.cost.cacheWrite = pricing.cache_write_per_m;
     existing.input = input;
     return existing;
   }
@@ -263,7 +265,7 @@ function transformApiModel(apiModel, existingModelsMap) {
       input: toNumber(pricing.input_per_m),
       output: toNumber(pricing.output_per_m),
       cacheRead: toNumber(pricing.cached_input_per_m),
-      cacheWrite: 0,
+      cacheWrite: toNumber(pricing.cache_write_per_m),
     },
     contextWindow,
     maxTokens,
@@ -392,8 +394,8 @@ function formatCost(cost) {
 
 function generateReadmeTable(models) {
   const lines = [
-    '| Model | Context | Vision | Reasoning | Input $/M | Cache Read $/M | Output $/M |',
-    '|-------|---------|--------|-----------|-----------|-----------------|------------|',
+    '| Model | Context | Vision | Reasoning | Input $/M | Cache Read $/M | Cache Write $/M | Output $/M |',
+    '|-------|---------|--------|-----------|-----------|-----------------|------------------|------------|',
   ];
 
   for (const model of models) {
@@ -402,9 +404,10 @@ function generateReadmeTable(models) {
     const reasoning = model.reasoning ? '✅' : '❌';
     const inputCost = formatCost(model.cost.input);
     const cacheReadCost = formatCost(model.cost.cacheRead);
+    const cacheWriteCost = formatCost(model.cost.cacheWrite);
     const outputCost = formatCost(model.cost.output);
 
-    lines.push(`| ${model.name} | ${context} | ${vision} | ${reasoning} | ${inputCost} | ${cacheReadCost} | ${outputCost} |`);
+    lines.push(`| ${model.name} | ${context} | ${vision} | ${reasoning} | ${inputCost} | ${cacheReadCost} | ${cacheWriteCost} | ${outputCost} |`);
   }
 
   return lines.join('\n');
@@ -429,9 +432,41 @@ function updateReadme(models) {
 // longer lists into deprecated-models.json (stamped with deprecatedAt) instead
 // of dropping them; the runtime appends them back so sessions and saved model
 // settings keep working, and after 14 days they are evicted permanently.
+//
+// The grace period exists for a model that is still ANSWERING while it leaves
+// the catalog. A retired one is not: Coral answers 404 with
+// `error.code == "model_retired"`, so keeping it registered only offers the
+// user a model that cannot reply, and /model still lists it for two weeks.
+// Each delisted model is probed once, and a retired one is dropped instead of
+// parked. Anything else — an unknown id, a key without access, a network
+// blip — keeps the grace period, because only `model_retired` is a statement
+// that the model is gone for everybody rather than missing for this key.
 const DEPRECATED_MODEL_TTL_MS = 14 * 24 * 60 * 60 * 1000;
+const RETIRED_PROBE_TIMEOUT_MS = 10000;
+
+/**
+ * Ask the gateway whether a model is retired. Returns true ONLY on an explicit
+ * 404 `model_retired`; every other outcome returns false, so a probe that
+ * cannot answer never shortens a grace period.
+ */
+async function isRetiredUpstream(id, apiKey) {
+  if (!apiKey) return false;
+  try {
+    const response = await fetch(CHAT_COMPLETIONS_URL, {
+      method: 'POST',
+      headers: { Authorization: `Bearer ${apiKey}`, 'Content-Type': 'application/json' },
+      body: JSON.stringify({ model: id, max_tokens: 1, messages: [{ role: 'user', content: 'ping' }] }),
+      signal: AbortSignal.timeout(RETIRED_PROBE_TIMEOUT_MS),
+    });
+    if (response.status !== 404) return false;
+    const body = await response.json();
+    return body?.error?.code === 'model_retired';
+  } catch {
+    return false;   // offline, timeout, unparseable body: keep the grace period
+  }
+}
 
-function updateDeprecatedModels(modelsJsonPath, newModels) {
+async function updateDeprecatedModels(modelsJsonPath, newModels, apiKey) {
   const deprecatedPath = path.join(path.dirname(modelsJsonPath), 'deprecated-models.json');
 
   let oldModels = [];
@@ -451,12 +486,25 @@ function updateDeprecatedModels(modelsJsonPath, newModels) {
   const added = [];
   const resurrected = [];
   const evicted = [];
-
-  for (const old of oldModels) {
-    if (old && old.id && !currentIds.has(old.id) && !deprecated[old.id]) {
-      deprecated[old.id] = { ...old, deprecatedAt: now };
-      added.push(old.id);
+  const retired = [];
+
+  const candidates = oldModels.filter((old) => old && old.id && !currentIds.has(old.id) && !deprecated[old.id]);
+  // Probe the delisted models plus anything already parked: a model can retire
+  // during its own grace period, which is exactly when it starts 404ing.
+  const parked = Object.keys(deprecated).filter((id) => !currentIds.has(id));
+  const toProbe = [...new Set([...candidates.map((m) => m.id), ...parked])];
+  const retiredIds = new Set(
+    (await Promise.all(toProbe.map(async (id) => ((await isRetiredUpstream(id, apiKey)) ? id : null))))
+      .filter((id) => id !== null)
+  );
+
+  for (const old of candidates) {
+    if (retiredIds.has(old.id)) {
+      retired.push(old.id);
+      continue;   // retired upstream: 404s, so a grace period buys nothing
     }
+    deprecated[old.id] = { ...old, deprecatedAt: now };
+    added.push(old.id);
   }
 
   for (const [id, entry] of Object.entries(deprecated)) {
@@ -465,6 +513,11 @@ function updateDeprecatedModels(modelsJsonPath, newModels) {
       resurrected.push(id);
       continue;
     }
+    if (retiredIds.has(id)) {
+      delete deprecated[id];
+      retired.push(id);
+      continue;
+    }
     const removedAt = Date.parse(entry && entry.deprecatedAt ? entry.deprecatedAt : '');
     if (Number.isNaN(removedAt) || Date.now() - removedAt > DEPRECATED_MODEL_TTL_MS) {
       delete deprecated[id];
@@ -472,9 +525,9 @@ function updateDeprecatedModels(modelsJsonPath, newModels) {
     }
   }
 
-  if (added.length > 0 || resurrected.length > 0 || evicted.length > 0) {
+  if (added.length > 0 || resurrected.length > 0 || evicted.length > 0 || retired.length > 0) {
     fs.writeFileSync(deprecatedPath, JSON.stringify(deprecated, null, 2) + '\n');
-    console.log('Updated deprecated-models.json ' + JSON.stringify({ added, resurrected, evicted }));
+    console.log('Updated deprecated-models.json ' + JSON.stringify({ added, resurrected, evicted, retired }));
   }
 }
 
@@ -517,7 +570,7 @@ async function main() {
 
     // Save models.json (pure API output, no patch/custom baked in)
     // Move delisted models to deprecated-models.json BEFORE models.json is overwritten
-    updateDeprecatedModels(MODELS_JSON_PATH, models);
+    await updateDeprecatedModels(MODELS_JSON_PATH, models, resolveApiKey());
     saveJson(MODELS_JSON_PATH, models);
 
     // Build the README model list: base → patch → custom. Grace-period models stay
@@ -549,4 +602,10 @@ async function main() {
   }
 }
 
-main();
+// Run only when invoked as a script, so the deprecation logic below can be
+// imported by tests without kicking off a sync.
+if (process.argv[1] && path.resolve(process.argv[1]) === path.resolve(fileURLToPath(import.meta.url))) {
+  main();
+}
+
+export { isRetiredUpstream, updateDeprecatedModels };